NVIDIA Isaac GR00T N1

NVIDIA's open humanoid-robot foundation model family, built on a fast-slow architecture and continually updated since its 2025 launch.

GR00T N1 is an open humanoid-robot foundation model (about 2.2 billion parameters) that NVIDIA released at GTC on March 18, 2025, with code under Apache 2.0 and weights under NVIDIA's open model license, both usable commercially. It uses a fast-slow, two-system design: System 2 is a vision-language model (Eagle-2 in N1) that looks at images and understands instructions; System 1 is a diffusion Transformer trained with flow matching that turns this into continuous action chunks, with each robot using its own state and action encoder/decoder. Data is organized as a “data pyramid”: web and human videos at the base, synthetic data from simulation and video models in the middle, and real-robot data at the top. Later releases followed: N1.5 (June 2025) froze the VLM and added a FLARE loss that lets it learn from human videos; N1.6 (December 2025) switched to a Cosmos-family VLM, doubled the action head, and moved to predicting relative actions; N1.7 (from April 2026), at about 3 billion parameters, switched its VLM to Cosmos-Reason2-2B and added pretraining on about 20,000 hours of EgoScale first-person human video.

ExampleNVIDIA provides GR00T-N1.7-3B on Hugging Face along with versions fine-tuned on LIBERO, DROID, and SimplerEnv; the README recommends a GPU with at least 16GB of memory for inference and 40GB or more for fine-tuning.

Also called

GR00T, GR00T N1 Series, GR00T N1.5, GR00T N1.6, GR00T N1.7, Isaac GR00T, Isaac-GR00T Repository

Related

Dual-System Architecture (System 1 / System 2) · Data Pyramid · Flow Matching · Diffusion Transformer · NVIDIA Generalist Embodied Agent Research Lab · EgoScale

Sources

GR00T N1: An Open Foundation Model for Generalist Humanoid Robots (arXiv 2503.14734)
NVIDIA/Isaac-GR00T GitHub 仓库 (Chinese)
GR00T N1.6 研究页（NVIDIA GEAR） (Chinese)

As of

2026-04